High-Bandwidth Flash or HBFD, the memory that offers more capacity than HBM, will not be used by NVIDIA; instead, Google is going to be a key customer.
NVIDIA To Keep on Using HBM, HBF Not Yet In Consideration, But Set For Sampling Later This Year
NAND DRAM has seen big adoption with the recent AI push. While primarily used for storage, such as SSDs, the upcoming flash memory technology could play a major role with HBF (High-Bandwidth Flash), the next generation NAND DRAM tech that bridges the gap between HBM and NAND Flash.
HBF is being co-developed by SanDisk and SK Hynix, and will use a similar architectural hierarchy as HBM, which is to stack multiple layers of NAND Flash on top of each other. Each layer will be connected using multiple TSVs (Through Silicon Vias), which will fuse all NAND packages into a singular stack. While HBM offers 32-64 GB capacities per stack right now, HBF will scale up to 4 TB capacities.
As for speeds, HBM is faster, but with architectural optimizations, HBF will be able to provide enough throughput for critical AI tasks. The HBF standard is ideal for inferencing workloads, which matter more today due to a surge in Agentic AI. The higher capacity also offsets some of the KV Cache constraints from the main compute chip.
While HBF sounds great, industry sources state that NVIDIA isn't planning to use the new DRAM tech any time soon, as it believes that eSSDs can address the constraints related to capacity and speeds. NVIDIA is reportedly working with Kioxia to produce PCIe Gen7 SSDs that are up to 100 times faster than standard designs.
As of right now, SK Hynix is leading the charge in HBF development with first samples planned for roll-out in the second half of this year.
Google is reportedly going to be a major customer for HBF as it plans to acquire the technology for its rapid AI expansion plans. Google's TPU ecosystem is growing at an accelerated pace, and they are scaling up their compute with several next-gen TPU solutions in the pipeline. Whether HBF becomes a prominent use case remains to be seen, but it does offer a huge application besides replacing HBM, and that's replacing standard DDR memory.
Servers have expanded LPDDR use as CPUs become the new constraint in AI. That has led to a big need for LPDDR5 and LPDDR5X memory, especially SOCAMM2. With HBF's multi-layered stacking approach, chip makers and AI ecosystem providers can not just reduce PCB space, but also add more capacities while keeping power usage low and retaining high bandwidth throughput.
About the author: A Software Engineer by training and a PC enthusiast by passion, Hassan Mujtaba serves as Wccftech's Senior Editor for hardware section. With years of experience in the industry, he specializes in deep-dive technical analysis of next-generation CPU and GPU architectures, motherboards, and cooling solutions. His work involves not only breaking news on upcoming technologies but also extensive hands-on reviews and benchmarking.
Follow Wccftech on Google to get more of our news coverage in your feeds.
